10 distributed databases that take you to the galaxy

Hello everyone , I'm Wukong .

This time, let's talk about databases in distributed scenarios . First of all, let's take a look at the differences and characteristics between relational and non relational databases .

One 、 Relational type vs Non relational

1.1 Relational type
1.1.1 What is relational ？

Relational databases refer to the use of relational models （ Two dimensional table model ） To organize a database of data , A data organization consisting of two-dimensional tables and their connections .

1.1.2 Common relational databases

Common relational database management system （ORDBMS）：Oracle、MySql、Microsoft SQL Server、SQLite、PostgreSQ、IBM DB2.

1.1.3 The advantage of relational

Using two-dimensional table structure is very close to the normal development logic .

Universal support SQL（ Structured query language ） sentence .

Rich integrity greatly reduces the problem of data redundancy and data inconsistency .

It can be used SQL It's a very complicated query between multiple tables ;

Relational databases provide support for transactions .

1.1.4 The disadvantages of relational type

（1） It stores row records .

Can't store array 、 Data in nested fields and other formats .

（2） It is inconvenient to extend the table structure .

An error will be reported if the operation does not exist , Adding columns requires execution SQL Statement . And you need to pay special attention to it , Because the table will be locked for a long time when it is updated , This can have a serious impact on the online environment .

（3） It takes up a lot of memory .

When a relational database performs statistics and other operations on a large number of data tables , It takes up a lot of memory , Because even if it only operates on one column , It also reads the entire line of data from the storage device into memory .

（4） The performance of full-text search is poor

Be similar to MySQL Relational database , Only use like Matching for whole table scanning , Efficiency is very low . Today, , There are many scenarios that need to support fuzzy matching , And it has to support efficient lookup . For example, query the log information containing keywords , Or query the product list according to a product keyword .

1.2 Non relational
1.2.1 What is non relational ？

NoSQL(NoSQL = Not Only SQL ), meaning " not only SQL".

A non relational database is not strictly a database , It should be a collection of data structured storage methods , It can be document or key value equivalence .

1.2.2 Common non relational databases

Key value database ：Redis、Memcached、Riak.

Columnar database ：Bigtable、HBase、Cassandra.

Document database ：MongoDB、CouchDB、MarkLogic.

Graphic database ：Neo4j、InfoGrid.

1.2.3 The advantage of non relational

Flexible format ： The format for storing data can be key,value form 、 Document form 、 Picture form and so on , Document form 、 Picture form and so on , Flexible use , Wide application scenarios , Relational databases only support basic types .
Fast ：NoSQL You can use hard disk or memory to store , The relational database can only use hard disk ;
High scalability ;
The cost is low ：nosql Database deployment is simple , It's basically open source software .

1.2.4 The disadvantages of non relational

Does not provide sql Support , The cost of learning and using is high ;
No transactions .MongoDB 4.0 Supported transaction .
The data structure is relatively complex , It's a little bit less in terms of complex queries .

Two 、 Distributed database

2.1 The definition of distributed database

There is no official definition of distributed database , It's just a conventional saying put forward by our technicians .

In the realm of databases , When the product is constantly evolving and gradually recognized by everyone , It will become a standard , For example, Microsoft's SQL Server database , Other databases like to contrast it , that SQL Server The database will become a standard .

But distributed database is also proposed in recent years , It's still relatively new , There is no reference . However, we can understand the distributed database through the experience of these big companies .

Distributed database is a database implemented with distributed architecture .

2.2 The advantages of distributed databases

Distributed has always been a topic of my research , Now many popular technologies use distributed architecture , For example, micro services 、 Message queue .

So why do we use distributed architecture ？ Simply speaking , It's using multiple computers （ machine ） To horizontally expand the performance of a single machine , Another important reason is distributed reliability , For example, multi machine backup 、 Disaster tolerance, etc .

Does the database also need to improve performance and ensure reliability ？ The answer is yes .

Which big factories are using distributed database ？

Every year 11, Ali likes it show A wave of trading record , Its distributed database OceanBase " . Big companies like Tencent 、 Bytes to beat 、 Meituan also began to use distributed databases , In addition, major banks have launched distributed databases .

So distributed database is a trend , If the business scenario requires high performance and high reliability , We can consider using the database under the distributed architecture .

2.3 Features of distributed database

First of all, let's take a look at two scenarios where databases are classified according to transaction types ：

online transaction （OLTP）

OLTP It's a transaction oriented process , The amount of data for a single transaction is small , But we have to give the results in a very short time , Typical scenes include shopping 、 Pay 、 Transfer, etc ;

Online analytics （OLAP）

OLAP Scenarios are usually operations based on large data sets , Typical scenarios include the generation of personal annual bills and corporate financial statements .

OLTP It is characterized by more writing and less reading 、 Low delay 、 High concurrency , So the database + Distributed in OLTP What are the characteristics of the scene ？

characteristic ：

It's powerful in the context of more writing and less reading .
Low latency response .
Support high concurrency .
Support mass storage .
high reliability .

3、 ... and 、10 A distributed database

3.1 PingCAP Of TiDB

Open source + Good community operation , Have a high popularity .

Definition ： It supports both online transaction processing and online analysis processing (Hybrid Transactional and Analytical Processing, HTAP） Integrated distributed database products , With horizontal expansion or reduction 、 Financial high availability 、 real time HTAP、 Cloud native distributed database 、 compatible MySQL 5.7 The protocol and MySQL Ecological and other important characteristics . The goal is to provide users with one-stop OLTP (Online Transactional Processing)、OLAP (Online Analytical Processing)、HTAP Solution .TiDB For high availability 、 Strong consensus requires higher requirements 、 Large data scale and other application scenarios .

TiDB Architecture diagram
TIDB Using a layered architecture , There are three kinds of characters ：

TIDB： As SQL engine .
TiKV： As the underlying distributed key value storage .
PD： Responsible for metadata management and global clock .
TiDB The derivatives of ：

Ti-Binlog、Ti-CDC Support data export .
Ti-Operator More convenient to implement container cloud deployment .
Chaos Mesh Support chaos Engineering .
shortcoming ： Global deployment is not supported , This is a cross regional large-scale cluster application TiDB Set up obstacles .

3.2 Google Of Spanner

Spanner It's developed by Google 、 Extensible 、 Many versions 、 Globally distributed 、 Synchronous replication database . It supports externally consistent distributed transactions .

F1 Mainly as SQL engine

Spanner Mainly responsible for transaction consistency 、 Replication mechanism 、 Scalable storage, etc .

Spanner The core processing module in the architecture is Spanserver,

Spanner The architecture of , come from Google The paper
Spanserver There are three parts to the core of our work ：

be based on Paxos Protocol data replication .Paxos The agreement can be seen in an article I wrote before ：《 Talk about distributed algorithms in Three Kingdoms , Comfortable, right ？》
be based on Tablet Slice management of .
be based on 2PC Transaction consistency management for .2PC The agreement can be seen in an article I wrote before ：《 Use Taijiquan to teach distributed theory , It is really comfortable ！》
2017 year ,F1 and Spanner It's split up , No longer a binding relationship . The principle is as follows ：

3.3 CockroachDB Cockroach database

CockroachDB ( Cockroach database ） It's a scalable 、 Support geographic processing 、 Data storage system supporting transaction processing .

Why is it called a cockroach ？

Because this database as long as the number of damaged nodes does not exceed half of the total number , So the cluster still works , Strong vitality .

Through distributed consistency algorithm instance to adjust to ensure consistency , It's chosen to use Raft Consistency algorithm . All consistent states exist in RocksDB in .

Cockroach It's a distributed one SQL database . The primary design goal is Extensibility , Strong consistency , Viability , Just like its name .Cockroach Our goal is to... Without human intervention , Tolerate disks with minimal interruption time , host , The rack even Data Center disaster .Cockroach The nodes of are peer-to-peer , One of the design goals is to have minimal configuration and no dependency , Deploy decentralized peer nodes . Chinese community address ：cockroachdb-cn.

CockroachDB Provides two different transaction features , Including snapshot isolation （snapshot isolation, abbreviation SI） Isolated from sequential snapshots （SSI） semantics , The latter is the default isolation level .

CockroachDB It's a distributed one K/V Data warehouse , Support ACID Business , Multi version value storage is its primary feature . The main design goals are global consistency and reliability , You can see this from the name of cockroaches . The cockroach database can handle disks 、 Physical machines 、 Minimum delay service interruption in case of rack or even data center failure ; The whole failure process does not need human intervention . Cockroach nodes are balanced , Its design goal is homogeneous deployment （ There's only one binary package ） And the minimum configuration .

CockroachDB and TiDB、YugabyteDB They all claim to be inspired Spanner, So it is often considered to be isomorphic products .CockroachDB and TiDB, It's often compared .

difference ：

CockroachDB Using standard P2P framework , As long as the number of damaged nodes is less than half of the total , So the cluster still works .
CockroachDB Support global deployment , Because it uses a hybrid logic clock （HLC）, So data consistency can be achieved on a global physical scale .
There are different management mechanisms of fragmentation .

3.4 YugabyteDB

Architecturally and CockroachDB There are many similarities , For example, supporting global deployment , Using mixed logic clock （HLC）, be based on Percolator Transaction model of , compatible PostgreSQL agreement .

Because of the high degree of similarity ,YugabyteDB And CockroachDB The competition is very fierce .

Yugabyte It adopts a two-tier architecture ： Query layer and storage layer . But the architecture is only logical , In the deployment structure , These two floors are located in TServer In progress . This and TiDB Different .

Yugabyte At the same time SQL and CQL Two kinds of API, among CQL Is compatible Cassandra It's a dialect grammar , The storage model corresponding to the document database ; and SQL API It's directly based on PostgresQL Demonic , It's compatible with PG grammar ,

Yugabyte It's the storage layer that matters . among TServer Responsible for the storage tablet, Every tablet Corresponding to one Raft Group, Distributed on three different nodes , To ensure high availability .Master Responsible for Metadata Management , except tablet Location information for , It also includes table structure and other information .Master It also depends on Raft High availability .

3.5 Ali OceanBase

OceanBase It is a financial level distributed relational database developed by ant group , From 2010 year .OceanBase With strong data consistency 、 High availability 、 High performance 、 Online expansion 、 Highly compatible SQL Standards and mainstream relational databases 、 Low cost and so on .

3.6 Tencent's TDSQL

TDSQL All of the nodes are used MySQL. It is a distributed cluster architecture （ Here's the picture ）, This kind of cluster architecture has high flexibility , Simplify the various Communication mechanism between nodes , It also simplifies the need for hardware . It doesn't just mean TDSQL A relational example of 、 branch Cloth example 、 Analytic instances can be deployed in the same cluster , It also means that even simple x86 The server , also You can build something similar to a minicomputer 、 Shared storage and other stable and reliable database .

3.7 Zte's GoldenDB

GoldenDB It is almost the largest distributed database in the domestic banking industry , and TDSQL Also select... On the data node MySQL, But the increase in global clock nodes makes it a standard PGXC framework .

3.8 Tencent's TBase

TBase It's Tencent's data platform team in open source PostgreSQL Enterprise level distribution based on R & D HTAP Database management system ：

• With high-performance and scalable distributed transaction capabilities , Support RC and RR Two isolation levels ;
• Through security 、 management 、 The system of separation of audit powers , Provide a comprehensive data security guarantee mechanism ;
• Support high performance partition table , The efficiency of data retrieval can be doubled ;
• SQL Aspect compatibility 2003 standard 、PostgreSQL Grammar and common use Oracle function & data type 、 Window functions, etc ;
• Provide data separation between large and small businesses 、 Efficient data management capabilities such as hot and cold data separation

There are three types of nodes in a cluster , They have different functions , Connect to a system through the Internet . The three node types are ：

• **Coordinator：** Coordinate nodes , Provide external interface , Responsible for data distribution and query planning , Multiple nodes are equally located , Each node provides the same database view ,CN Global metadata of the storage system .
• **Datanode：** Process and store metadata related to this node , Each node also stores a slice of data . Functionally ,DN The node is responsible for completing the execution request distributed by the execution coordination node .
• GTM: Global transaction manager (Global transaction manager.), Responsible for the management of cluster transaction information , At the same time, manage the global objects of the cluster , Example sequence , besides GTM There are no other functions available on .

3.9 VoltDB

VoltDB The profile provided on the official website ：VoltDB It's the fastest in memory database in the world , It inherits the strong consistency requirement of traditional relational database , It also provides the ability to deploy and distribute on the Internet cloud Horizontal expansion capability of database .VoltDB By saving all the databases in memory , Eliminates massive data and log disk access operations , In a single threaded way , Eliminates disk lock and record lock ; Through database fragmentation Technology , Let the database support high concurrent requests ; Support horizontal expansion of database through distributed cluster . Its query speed is as fast as that of traditional database 100 More than times .

2019 It was officially closed in , Into pure commercial products . At the same time ,VoltDB There is no complete service support system in China , This affects its promotion to a large extent .

3.10 Giant fir SequoiaDB

SequoiaDB Jushan database is an open source financial level distributed relational database , It mainly provides high performance for high concurrent online transaction scenarios 、 Reliable, stable and unlimited horizontal expansion of database services .

Users can go to SequoiaDB Create multiple types of database instances in Jushan database , In order to meet the needs of different upper applications .

SequoiaDB Jushan database supports MySQL、PostgreSQL、SparkSQL and MariaDB Four relational database instances 、 class MongoDB Of JSON Document class database instance 、 as well as S3 Object storage and POSIX Unstructured data instance of file system .

SequoiaDB Jushan database storage engine adopts distributed architecture . Each node in the cluster is an independent process , Between nodes TCP/IP Protocol to communicate .

The same operating system can deploy multiple nodes , Different ports are used to distinguish nodes .

Okay , For distributed databases , If you also have experience with distributed databases , Welcome to leave a message ～

Reference material ：

https://docs.pingcap.com/zh/tidb/stable
http://vldb.org/pvldb/vol11/p1835-samwel.pdf
http://ericfu.me/yugabyte-db-introduction/
https://blog.csdn.net/duan_zhihua/article/details/88549173
https://www.voltdb.com/
https://www.zhihu.com/question/24225007/answer/1707736658
https://www.zhihu.com/question/24225007/answer/1326259742
https://time.geekbang.org/column/article/296558